acceptodds
Under review as a conference paper at ICLR 2027

Beyond Distribution Matching: Kinetic Constraints for Unpaired Multi-Omics Alignment

Abstract

Unpaired static multi-omics data does not clearly answer which cells or molecular states correspond across modalities as distinct multi-omics mappings can produce the same target distribution with potentially biologically impossible states. We introduce Kinetics induced Optimal Transport (KOT), a framework that combines distribution matching with first-order constraints on the differential of the cross-modal mapping itself. KOT pushes an observed source vector field through the mappings own Jacobian, requiring the induced target dynamics to be in agreement with a mechanistic target law, without the need for paired cells or observed target velocities. We show theoretically that this constraint reduces distribution-preserving alignments, that ignore the dynamics of single-cell biology, resulting in identifiability up to shared dynamical symmetries. For RNA–protein transport, biochemical anchors resolve a global kinetic-scale ambiguity. On CITE-seq datasets, KOT reaches competitive performances for cell-level RNA–protein alignment in an unpaired setting. Moreover, a mapping trained only on wild-type cells also transfers to out-of-distribution CRISPR perturbations. Mechanistic ablations show that state correspondence, internal kinetic consistency, and biological dynamics are distinct properties of KOT. Lastly, a chromatin–RNA extension exposes the boundaries of KOT. First-order constraints effect the learned differential, but cannot provide source-field-specific ambiguity resolution when the available chromatin directional field does not contain enough information. These experiments further establish first-order dynamical consistency as an inductive bias for learning unpaired cross-modal correspondences beyond static distribution matching. Code is provided in https://anonymous.4open.science/r/KOT/README.md

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Optimal Transport, Physics Informed Neural Network, Single-Cell Multi-Omics, Unpaired Cross-Modal Integration

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_jKFBOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces KOT, which uses source-modality dynamics and mechanistic relationships in the target modality to learn unpaired cross-modal mappings. The RNA–protein alignment and CRISPR protein-effect prediction experiments show promise. I find the direction meaningful, but it remains unclear which information drives the improvements and how the additional requirements affect the method’s practical use.

Critical strengths

1.The idea of incorporating dynamics into unpaired learning is clear and supported by controlled synthetic experiments.

2.The method achieves good alignment results on two CITE-seq datasets. A model trained only on non-targeting cells also retains predictive signal for protein effects under unseen perturbations.

3.The ablations are extensive. The authors’ discussion of the differences between static alignment, kinetic consistency, and perturbation prediction helps clarify the method’s capabilities and limitations.

Critical weaknesses

1.The contribution of dynamics to static alignment needs clarification. Shuffling, reversing, or even zeroing the velocity leaves alignment performance close to that of the full model, whereas disrupting the gene–protein correspondence substantially worsens it. This suggests that the static alignment gains may come largely from the structural prior. Comparing against a model with the entire kinetic loss removed does not sufficiently separate these contributions.

2.The perturbation results are encouraging, but the conclusions should remain limited. The model receives experimentally observed post-perturbation RNA, and the evaluation covers population-average effects for a small protein panel in one dataset. The advantage over a simple cognate-mRNA reference is also modest. These results support some transfer capability, but do not yet establish broadly accurate recovery of intervention responses.

3.The practical benefit of the unpaired setting is not yet fully demonstrated. The method removes the need for cell-level paired supervision, but still requires data suitable for velocity estimation, gene–protein relationships, and, in some experiments, half-life references. The paper could explain more concretely when these inputs are easier to obtain than paired measurements. The current benchmarks hide correspondence in originally paired datasets and therefore do not directly test the full workflow on independently collected populations. Runtime measurements on mixed hardware provide only a rough reference, and the cost of velocity preprocessing also needs to be considered.

Questions for authors

1.With the same gene–protein prior retained, how much additional benefit comes from using the correct dynamics? We would appreciate a clearer comparison that isolates the contribution of the dynamics model from that of the prior.

2.Can the perturbation results be reproduced across more proteins or datasets? Comparisons with methods given access to the same information would also help readers assess how broadly the reported improvements generalize.

3.Do preprocessing and model selection also avoid cell-level pairing information? Please clarify how RNA size-factor normalization is performed in Papalexi and whether FOSCTTM is used during model development, since these details affect the fairness of the comparison.

4.What practical setting do the authors consider best suited to KOT? Could you also report the total runtime, including velocity preprocessing, and peak memory usage on consistent hardware? This would help readers better assess the computational cost of using the method.

Reviewer_4xzcOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduces KOT (Kinetics induced Optimal transport) to improve unpaired multi-omics alignment. THe key idea is to introduce a constraint baised on the Jacobian of a learned mapping. The authors also provide some theoretical discussions which focus on residual symmetries. The experiments are based on synthetic, CITE-seq, and CRISPR data.

Critical strengths
  • The constraint based on the Jacobian is a nice inductive bias, and the authors provide a symmetry analysis and a synthetic setup that tests it.
  • In the experiments based on CITE-seq and CRISPR, the method outperforms the unpaired controls.
Critical weaknesses
  • It seems that the gain from the alignment seems to come more from the kinetic term and transcript–protein correspondence than from the observed direction of cellular dynamics (in Tab 3, we see that shuffling / reversing / zeroing the source velocity does not degrade performance that much compared to the full KOT).
  • On CRISPR, we see in Tab 2 that perturbing the transcript–protein correspondence slightly improves all three reported metrics (Table 2). hence there is not a clear gene-specific mechanistic interpretation of that result.
Questions for authors

Can a regularizer which uses the same transcript–protein links and model capacity but without cell-specific velocity achieve similar gains as in the CITE-seq exp? Could you provide intuition on why perturbing the transcript–protein correspondence improve CRISPR transfer? Also, how stable is that result across perturbation groups?

Reviewer_ahrQOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper tackles the alignment of unpaired single-cell multi-omics data, where matching marginal distributions does not identify the cell-level correspondence. The authors propose Kinetics-induced Optimal Transport (KOT), which learns a map from RNA to protein states using a Sinkhorn divergence between the mapped and observed populations, together with a kinetic term. This term requires the Jacobian-vector product of the map with the RNA velocity to agree with a translation-degradation law whose parameters are learned jointly with the map. Experiments cover a synthetic branching benchmark, two CITE-seq datasets, prediction of protein effects of unseen CRISPR perturbations, and an extension to chromatin-RNA alignment.

Critical strengths
  • The paper studies the important problem of unpaired multi-omics alignment.

  • The paper includes ablations and honestly reports negative results, such as the chromatin-RNA experiment.

Critical weaknesses
  • The paper is not finished in its current form, making it hard to read. Large parts of the text are generic and repetitive. The paper contains many typos, LaTeX artifacts, unfinished sentences, and wrong cross-references. Several methods that appear in the tables are missing from the bibliography, including MaxFuse (Chen et al., 2024), totalVI (Gayoso et al., 2021), and uniPort (Cao et al., 2022), and the PBMC dataset appears to be misattributed. The related work also omits relevant unpaired alignment methods such as scTopoGAN (Singh et al., 2023), InfoOT (Chuang et al., 2023), and DDMEC (Bounoua et al., 2025).

  • The empirical evidence does not support the central claim. On real data, shuffling, reversing, or zeroing the velocity leaves the alignment nearly unchanged, while permuting the transcript-protein correspondence destroys it. This indicates that the gains come from the link between each protein and its cognate transcript, which existing methods already use.

Cao, K., Gong, Q., Hong, Y., & Wan, L. (2022). A unified computational framework for single-cell data integration with optimal transport. Nature Communications, 13(1), 7419.

Chen, S., Zhu, B., Huang, S., Hickey, J. W., Lin, K. Z., Snyder, M., ... & Ma, Z. (2024). Integration of spatial and single-cell data across modalities with weakly linked features. Nature Biotechnology, 42(7), 1096-1106.

Chuang, C.-Y., Jegelka, S., & Alvarez-Melis, D. (2023). InfoOT: Information maximizing optimal transport. In International Conference on Machine Learning (pp. 6228-6242). PMLR.

Gayoso, A., Steier, Z., Lopez, R., Regier, J., Nazor, K. L., Streets, A., & Yosef, N. (2021). Joint probabilistic modeling of single-cell multi-omic data with totalVI. Nature Methods, 18(3), 272-282.

Bounoua, M., Franzese, G., & Michiardi, P. (2025). Learning to match unpaired data with minimum entropy coupling. In International Conference on Machine Learning. PMLR.

Singh, A., Biharie, K., Reinders, M. J., Mahfouz, A., & Abdelaal, T. (2023). scTopoGAN: unsupervised manifold alignment of single-cell data. Bioinformatics Advances, 3(1), vbad171.

Questions for authors
  1. How does KOT compare with a Sinkhorn objective combined with a linked-feature term, without velocity?

  2. Which model produces the synthetic Branch result in Table 1, and does it use simulator kinetics?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.